Companies are ordering staff to lean on AI, then handing the credit to the machine. Researchers call it the AI penalty, and workers say it is costing them promotions and raises.
Aubrey spent more than a year on a project to speed up a costly medical manufacturing process. When she finished, her manager asked her to present it to senior leadership as if Claude, the AI chatbot, had done the work. She compromised, talking up the bot while making clear she had carried the load.
Her manager cut in anyway and told the room she had built the whole thing in a minute with AI. Weeks later her annual review came back lukewarm, and her boss admitted the moment had counted against her.
She is not alone, as Business Insider’s Shubham Agarwal reports. Deepak, an IT developer at a Fortune 500 firm in India, began openly crediting the coding agents he used, in the name of transparency. Before long his managers assumed all his good work was the machine’s, and he thinks it stalled a promotion.
Many white-collar workers are now caught between bosses who demand more AI and bosses who dock its users. So they have started hiding how much they lean on it.
The AI penalty #
The instinct to hide it is rational. Christoph Riedl, a professor at Northeastern University, ran a meta-analysis of 13 studies across many jobs and found a clear pattern. Managers consistently marked work down once staff admitted a chatbot had helped, assuming the machine did most of it. Riedl calls this the “AI penalty”.
The main way to dodge it, he found, is to keep control of the core task and spell out exactly what you did. That is hard when bosses roll out ever-blunter ways to watch AI use, in a job market already rattled by automation.
You cannot measure creativity in tokens #
The bluntest measure is the token, the basic unit an AI model processes.
Counting tokens tells a manager how often someone prompted a chatbot and how much text moved back and forth. It reveals nothing about what the AI actually contributed. Staff soon learned to game it, firing off pointless questions to look like power users, so firms are now reining that in.
Last month Amazon scrapped an internal leaderboard that ranked token use. “Please don’t use AI just for the sake of using AI”, a senior vice-president, Dave Treadwell, told a company meeting.
Even finer tools mislead. Coding assistants such as Claude Code stamp a co-author line on the code they help write. They do not flag which lines were theirs or how much the human shaped them.
“If AI use is disclosed without specific details about how it was used, the manager’s default assumption seems to be that it was used in a way that reduces agency”, Riedl says. In plain terms, bosses assume the bot led. So the how, he adds, matters enormously.
The tools trying to fix it #
Some researchers are trying to pin down the split between human and machine. Graham Neubig, a computer scientist at Carnegie Mellon University, cofounded OpenHands. The open-source coding platform footnotes any line an AI wrote, so reviewers know to scrutinise it.
A team at IBM went further with an AI Attribution Toolkit, modelled on the system scientists use to credit each author of a paper. It lets people log how much a chatbot generated and what a human checked, then produces an attribution statement.
High-level nods to AI are not enough, says Jessica He, one of its designers.
Transparency’s paradox #
The deeper problem is social.
Several studies show that disclosing AI, even honestly, makes colleagues trust you less and read you as lazy. Oliver Schilke, a professor at the University of Arizona, found the same in his own work. He calls it a paradox: the people who do the honest thing pay for it. He wants shared rules on AI credit rather than each worker guessing.
Thomas Prommer, an engineering executive at Adidas, watched mandatory attribution backfire. His engineers stopped reaching for AI at all, he says, “because they didn’t want their best contributions footnoted as ‘cowritten by Claude’”. What worked instead was crediting the outcome, not the tool.
The stakes run past a missed raise.
Earlier this year Amazon was found to have blamed staff, and laid them off, for an AI agent’s mistakes. “The praise goes to AI, but going through its content is our responsibility”, Deepak says. Alessio Artuffo, chief executive of the learning platform Docebo, argues that simple attribution is the wrong frame.
The real question is not how work was made but whether the person behind it can defend and fix it. If firms keep punishing the people who use AI honestly, he warns, they will get more output and less ownership. That, he says, is “capability regression dressed up as efficiency”.
Get the TNW newsletter #
Get the most important tech news in your inbox each week.